Performance Evaluation of Hardhat Image Classification Using Deep Learning and Grad-CAM

Safety helmets function is to protect the worker's head from hazards in the work environment. Nevertheless, there are still many workers continue to neglect wearing safety helmet when they are in the workplace. Lack of supervision and the large size of the work location are also some of the obstacles that exist in large-scale companies which have the potential to increase the number of accidents. This research adopted Convolutional Neural Network (CNN) algorithm to classify humans wearing hardhat and not wearing hardhat based on digital images. Many safety helmet detection systems have been built using various Convolutional Neural Network (CNN) algorithms, but there is still limited information regarding the performance comparison of the architectures used. From these problems, this research was conducted to evaluate the performance of architectures that had been built previously, such as YOLO (You Only Look Once), EfficientNet, ResNet, and VGG. This research uses a dataset containing 1800 images. Divided into two classes, namely with hardhat and without hardhat, each with 900 images. Following this research, it was discovered that the constructed model was able to classify the image of safety helmets. The study's findings demonstrated that YOLOv5x was more accurate than other architectures, namely 99.5%, followed by YOLOv5l at 99%, YOLOv5m at 99%, EfficientNet at 98.9%, YOLOv5s at 98.5%, YOLOv5n at 98.5%, ResNet50 at 96.1 %, and VGGII at 95.5%. This research also conducted grad-CAM as the Explainable Artificial Intelligence (AI) which enlighten on why the CNN architectures made the prediction.